Most edge orchestrators must answer one question reliably: when can a compact local model safely handle a request, and when must it triage to a larger frontier model? This dataset encodes that decision boundary as structured training signals rather than a single label, letting small routing models learn intermediate features (domain, complexity, task type) before predicting a route.
What Sets It Apart
- Feature-rich rows: each example pairs a user prompt with a full reference solution that preserves intermediate reasoning and code blocks, plus explicit attributes (complexity_score, primary_domain, coding_task, math_task, requires_reasoning) so a model can learn to predict intermediate features before routing.
- Operationally focused rule: a deterministic override maps complexity>=3 or coding/math tasks to “big model”, creating a clean operational boundary that reduces frontier-model self-reflection bias and yields consistent routing targets during fine-tuning.
- Edge-first design: curated to train SLMs (target 50M–1.5B) with a recommended multi-task sequence-generation objective that has demonstrably higher routing accuracy than naive single-token classification.
- Companion router model: the dataset ships alongside a reference 51M routing model intended as a starting point for deployment experiments.
Who It's For and Trade-offs
Great fit if you are building on-device/orchestrator routing for constrained hardware, training compact routers to triage code/math/complex reasoning away from local models, or evaluating multi-step supervision strategies. Look elsewhere if you need large-scale real-user logs (this dataset is ~1k examples and synthetically seeded) or if you require diverse multilingual coverage—entries are English-focused and generated from seeded corpora, so synthetic biases and limited size are real constraints.
License: MIT. Size and provenance: ~992 training examples, created with seeded generation from larger reasoning and instruction corpora and programmatic overrides to enforce routing policy.
